In this article, the authors provide a way relevant work and dataset for the field, which promises significant future use by a wide number of different stakeholders. That also implies an important responsibility to ensure the precision and accuracy.

Getting to the content, I believe that reviewer#1's comments (after 1st iteration) about uncertainties of the ARGO-O2s should be addressed before publication, I copy them below:

"1. The ARGO-O2 is an evolving technology with variable accuracy and uncertainties even in the delayed mode dataset. Can we confirm whether or not GOBAI-O2 uses only optode sensors calibrated by the known methods (either climatology or in-air O2 measurements)? If other types of sensors are used, it should be stated.
2. The uncertainty range should be re-considered. Of the O(900) floats that passed the QC step, there is a diversity of sensors and calibration methods. Only relatively new ARGO-O2 profiles are calibrated with in-air O2 measurement, and the older data (essentially all floats deployed 2015) had to be calibrated using climatology. Because GOBAI-O2 blends all kinds of O2 sensors, and uncertainties should be re-assessed in section 2.5. My suggestion would be no less than 3% for measurement uncertainty considering the climatological calibration from older O2 sensor profiles.
3. In section 2.5, the combined measurement uncertainty is stated as 1.5%. I believe this is an average of 1% in GLODAPv2 and 2% in ARGO-O2. As stated above, ARGO-O2 should have 3% uncertainty at least, and the combined uncertainty should consider the number of profiles from each dataset. My reading is that the authors used O(21k) profiles from GLODAPv2 and O(133k) profiles from ARGO-O2. Therefore, measurement uncertainty should be dominated by the ARGO-O2, and due to the blend of unaccounted sensor/calibration type, it should be 3% in my opinion.
4. The issue of sensor response time is briefly discussed in section 2.5. The community has not yet implemented the response time correction (RTC) of optode sensors. In section 2.5 the authors stated that 2% uncertainty of ARGO-O2 is still optimistic, but I don’t think this statement adequately describe the problem. The lack of RTC in ARGO-O2 data not only increases the measurement uncertainty but also causes systemic bias in the vicinity of oxycline. This is a larger problem beyond the scope of this paper, however, it should be discussed and the data users must be warned. Much stronger statement of caution should be provided, perhaps in the discussion/conclusion. Current ARGO-O2 dataset can include significant bias and uncertainty in the oxycline region beyond the level that is characterized in the uncertainty analysis. Simply adding a constant offset or blending GLODAPv2 data using machine learning will NOT correct this issue. Progress in this area will be much needed for future research."

And just as a minor comment, I'ld have preferred the usage of ML term instead of AI in the work, as the latter tends to be used for smart systems with take autonomous decisions, but as the term definition is very ambiguous and broad (and includes ML), I've no objections.